feat(interval): add feedback-guided search policy

[kim-em]

Summary

• learn contractor productivity from engine-authenticated fact improvements and deterministic logical work
• give untried actions an optimism bonus, decay repeated fixed points, and retain an explicit finite-frontier fairness tier
• retain deterministic bounded MRU observations keyed by stable action sites and authenticated snapshots
• keep scoring independent of proof evidence, fact meaning, operation names, and package keys
• exercise synthetic and live arbitrary-function choices

Trust boundary

Policy scores select already-authenticated work only. Engine admission and ordinary kernel proof replay remain unchanged. No native_decide, axioms, or sorries.

Checks

• focused adaptive/staged-policy, exact-branch, subtraction, nested-branch, and complete LogTables conformance build
• trust surface, factor freshness, PNT source inventory, diff, and banned-proof checks